>I tried rolling in the recent pmap.c/trap.c changes, (from pmap.c version
>1.44) which move the pv attributes into the pv_table, and am seeing 
>processes getting corrupted and dying.

>My crude test case is building gnu make, and I see gcc die with
>syntax errors.

>I haven't seen this in a simulated environment (which lacks the 
>cache coherency issues) but the simulator also has much more memory,
>so it's not clearly a caching problem. 

>Anyone else seen similar behavior?  This is with the old VM system, but
>I believe similar behavior occurs with UVM.  (Just haven't rebuilt and 
>tested lately.)

hi Castor,

I havent seen this with UVM. I've rebuilt the world twice with UVM
using egcs 1.1.1, and I can (at last) run mdsetimage without kludging
it to do msync(). I was thinking of kicking out a snapshot with UVM.
acutally.

IIANM, all but the most-recent bunch of changes (the ones which add a
PV_REFERENCED bit as well as PV_MODIFIED) were also tested under the
old Mach VM.  Perhaps the PV_REFERENCEDthose changes need to be inside
"#if defined(VM)" as well.